Hypertun-SR
Reference
This repo is an implementation of paper: High-Performance and Tunable Stereo Reconstruction
The author of this paper has not provided source code. Johann Diep provides an unofficial implementment: Original repo. This repo is mostly based on Johann Diep's work and performs several slight changes.
Thanks Sudeep Pillai, Johann Diep and everyone who have make effect to this algorithm!
Requirements
-
OpenCV 2.4.13+ ( We have modified exFAST's interface to fit OpenCV 3.X. Additionally, we currently NOT recommend replacing exFAST/FAST with ORB feature detector in ORB_SLAM2 after testing. )
-
boost (tested with 1.58.0.1)
The pipeline also uses the exFAST library, which is included in ThirdParty/sparsestereo and the Triangle library, which is included in ThirdParty/triangulation.
Datasets
Two KITTI datasets are used which should be contained in folders as:
-
Hypertun-SR/data/data_stereo_flow (http://www.cvlibs.net/datasets/kitti/eval_stereo_flow.php?benchmark=stereo)
-
Hypertun-SR/data/data_odometry_color (http://www.cvlibs.net/datasets/kitti/eval_odometry.php, only the folder 00)
These two datasets can be toggled by changing the variable int DATASET in the main.cpp.
Compiling and running
The code was already build on Ubuntu 16.04 using cmake. To run the pipeline use the executable located at test/bin/main.out.
To build the code again, navigate into the project source folder and use:
cmake ..
and then
make
Efficiency Improvements
The efficiency of the original repo has great room for improvement. We implement several changes to accelerate the pipeline.
The original repo running at 16Hz with 2 iterations (in Report), we have improved it up to around 30Hz, which is close to the data provided by original paper (34Hz).
However, since we focus more on the efficiency, the accuracy is still lower than the original paper present, same as original repo.
-
Extracting high gradient pixels and preserving the coordinates simultaneously.
-
Using Scanline Algorithm to allocate triangulates' indexs to all pixels.
Accuracy Improvements (TODO)
This part is still not so prefect, currently I can only improve the accuracy mainly by finetuning the parameters and with the sacrifice of efficiency.
Currently the frequency has dropped from 30+Hz (on my desktop) to 26Hz (on my laptop), 34Hz in Report and the accuracy (disp<3pix) has raised from 75.8% to 85.1%, 90.2% in Report. For more info please refer to CHANGELOG
Currently this repo's results are as follow (ON MY LAPTOP):
